It’s not super clear what is the boundary between normal maps and HD maps, but HD maps have higher detail. Normally HD maps are used to augment perception sensors, ie the car can use the map to see more/better than the perception sensors can see. To be able to this information it is often required to localize oneself in the map using a localization algorithm. Google have very high definition HD Maps made using Lidar pointclouds, once they localized themselves in these driving becomes very straight forward unless you hit an edge case. Elon Musk claims that this can make you stuck in a local optimum during development phase and that as humans can learn to drive without HD maps computers should be able to learn this also.
Humans learn to drive without pinging radio pulses from their nose so why has Tesla been stuck in the local optimum of relying on a crappy radar, mainly to produce phantom braking events, for the last 4 years?
Karpathy stated in Apr. 2019 that it's belatedly being used as a ground truth input to train vision NNs to accurately estimate depth to vehicles ahead, but even then this is still a dangerous crutch, since when moving at highway speed it cannot train NNs to recognise stationary obstacles which are essentially invisible to it.
This foreseeably leads [ given the known vagaries of human nature ] to the classic Tesla phenomenon of AP tilting at full cruising speed into large unrecognised stationary objects, the most recent example being a
M3 ploughing into the roof of an overturned truck on a motorway in Taiwan.
In which case vision is presumably reporting "
huge sudden discontinuity of motorway road surface dead ahead (i.e. end of drivable space)", radar says "
fuggeddaboudit" and maps are apparently not consulted but (even non-HD maps) would return "
we really should be seeing more motorway around here" -- nevertheless, the resulting decision from these conflicting inputs is in fact "
föök it, full steam ahead, the hyperalert human will catch any glitch without even a warning tone, or, if not, that *sugar*'s probably insured!"
At this juncture I'd prefer the local optimum of sparing the life/limbs of any less than totally attentive human driver, at minimum by bleating a tone to indicate there is some weirdness ahead or better yet by applying full emergency braking.
Why is that so difficult for Tesla?
Why does map+vision not override the known unreliable radar?
If an earthquake in California opens a 100 foot gap in a highway flyover, would all the Teslas on AP by default just cram into the chasm?
If a 40 foot container falls off the back of a truck, perpendicular across a two lane highway, blocking it completely but just so happens to have a graphic painted on the side depicting a wide open road stretching off into an idyllic countryside and the lane lines perfectly match up
(the Roadrunner painting-a-tunnel-on-cliff scenario) -- how does AP/FSD @80mph deal with this? At the moment it certainly looks like full speed collision is what will happen.
At an individual level being stuck in a local optimum is preferable to being plastered over a pre-collapsed crash-attenuator @70mph after AP follows a fake line across a gore area (case of Walter Huang in CA), which HD maps would make impossible.
